https://arxiv.org/abs/1706.07068
It's a little different from GAN. That is, the classifier also learns style. And the paper that the generator will be more creative than the conventional GAN by learning the generation so that the style is also different. As an art dissertation, it turns out that artists often aim for style breaks. Generate generators to maximize style deviations while minimizing deviations from the art distribution
A scholar who is a requirement of the creative system states that one imagination, two skills (quality), and three ability to evaluate unique creations are required, but all three are satisfied. One of the main features of the proposed system is to learn about the history of art in the process of creating art. However, there is no semantic understanding of the art behind the concept of style. I know nothing about the subject, the explicit model of the element, the principle of art. The learning here is based solely on the concept of exposure to art and style. In that sense, the system has the ability to continually learn from new art and adapt generations based on what it learns.
It is based on the theory proposed by the old DE Berlyne (1924-1976). He emphasized that the most important arousal-enhancing properties of aesthetics are novelty, surprise, complexity, ambiguity, and mystery. And if one artist keeps making works, he will get used to it, so I will do my best with this system to avoid it. Also, the stimulus is neither too strong nor too weak, so I control it. There are GAN extensions that make it easy to generate images based on categories (eg [18]) or captions (eg [19]). ). By providing training on such labels, you can think of GANs that can be designed and trained to produce images of different art styles or different art genres.
Doesn't require humans for feedback. There was googleDream, but it's too vague, it's not abstract art, and it's said to be computer-generated. It's too unrecognizable. Do your best with moderate ambiguity
The discriminator also returns to the generator whether it is art or not, like a normal GAN. It returns to the generator a value of how much it can be classified into which style. The classifier has access to a large number of art related to style labels (Renaissance, Baroque, Impressionism, Expressionism, etc.) and uses them to learn the distinction between styles. The generator strives to create what is art and to confuse the classification as much as possible.
Of the various mechanisms of arousal, one that is particularly important and relevant to art is the characteristic of external stimulus patterns [3]. Martindale emphasized the importance of familiarity in deriving art production systems. [15] Wundt curve (curve that measures the degree of arousal?) He commented on the images generated by Google DeepDream [16]: "Most of them look like a mandala in a dorm room, or a digital psychedelia that appears to be on the cover of a book by Terence McKenna." Others commented, "Dazzling, drugs, and creepy." 4. This negative reaction can be explained as a result of excessive arousal, resulting in negative pleasure according to the Wundt curve.
Recommended Posts